---
title: "Claude Code Worktrees as an AI Company: How Autonomous Is It?"
date: "2026-09-17"
video_id: "7686612195530050829"
tiktok_url: "https://www.tiktok.com/@apop/video/7686612195530050829"
creator: "@6775612853199406086"
category: "AI"
tags: ["claude code", "git worktrees", "subagents", "cross-session messaging", "commercial real estate"]
excerpt: "A commercial real estate CRM built as an AI org chart in Claude Code runs on four worktree sessions, and a human still carries every prompt between them."
---
A TikTok from @apop (Alex, who works in industrial real estate) shows a commercial real estate CRM being built by what he calls "a fully autonomous AI company" in Claude Code. The setup is four Claude Code sessions, each in its own git worktree, with subagents doing the work inside each one. It is a sensible structure, but it is not autonomous: Alex copies prompts from one session to the next by hand, and his own org chart says he approves every new branch.

## How the setup works

The org chart on his screen, for a company he calls AYXCO, describes itself in one line: "worktree = sidebar group = session title." Four sessions sit at the top: `ceo` for questions, plans and reviews, `coo` for records data, `cto` for the product and infrastructure, and `cro` for revenue.

Under them sit 18 desk definitions. The `coo` branch has nine desks (`book`, `sales`, `clerk`, `tax`, `entities`, `owners`, `zoning`, `edgar` and `signals`), `cto` has five (`ui`, `ops`, `automations`, `ghl` and `skiptrace`), and `cro` has three (`prospecting`, `clients` and `marketing`). A separate auditor, "a fresh-context reviewer", sits outside the chart and "reports gaps against the stated requirement."

A root `CLAUDE.md` that every branch loads acts as the constitution. Its rules include "simplicity · permanent fixes only · measured not asserted · public record only · money and risk are the board's," and a work loop of "explore → plan only if more than a sentence → build small → verify with a check that can fail → auditor → land."

## What Claude Code features does it use?

Worktrees. [Anthropic's worktrees page](https://code.claude.com/docs/en/worktrees) says that "in the desktop app, select the **worktree** option when you start a session to give it its own worktree," which matches the checked box in the video. Worktrees are "created under `.claude/worktrees/<name>/` at your repository root, on a new branch named `worktree-<name>`," and the page states the reason: "Running each Claude Code session in its own worktree means edits in one session never touch files in another."

Sidebar groups. This part of the chart overstates the product. The [desktop app page](https://code.claude.com/docs/en/desktop) says you can "group sessions by project." Grouping keys off the project, so the `ceo`, `coo`, `cto` and `cro` labels are session names Alex typed and keeps consistent himself.

Subagents. The desks Alex calls "employees" match Claude Code's [subagents](https://code.claude.com/docs/en/sub-agents). "Each subagent runs in its own context window with a custom system prompt, specific tool access, and independent permissions," and a subagent "can spawn subagents of its own, up to three layers below the main conversation." Each subagent can pick its model through a frontmatter `model` field that accepts "`sonnet`, `opus`, `haiku`, or `fable`."

## Does Fable use more tokens than Opus?

Alex says the CTO session "works off of fable which uses more tokens" while the workers run on "opus which is using less tokens." The model on screen is "Fable 5.1" at "High" effort.

The difference he is describing is mostly price. [Anthropic's models overview](https://platform.claude.com/docs/en/about-claude/models/overview) lists Claude Fable 5.1 at "$10 / input MTok, $50 / output MTok" and Claude Opus 5.5 at "$4 / input MTok, $20 / output MTok." Fable costs 2.5 times as much per token on both input and output.

There is some truth in the token framing. The same table lists Fable 5.1's default effort as `high` and Opus 5.5's as `medium`. The [cost page](https://code.claude.com/docs/en/costs) says "thinking tokens are billed as output tokens" and that "you can't turn off thinking on Opus 5.5, Sonnet 5.5, or the Fable models." A Fable session at high effort will usually produce more thinking tokens per turn, so it uses more tokens because of the effort setting.

Anthropic's guidance points lower for workers. The cost page says "Sonnet handles most coding tasks well and costs less than Opus," and for delegated work suggests you "choose a smaller model for a subagent." Sonnet 5.5 costs $2/$10 per million tokens.

## Is the system fully autonomous?

No, and the chart on the same screen says so. Its subtitle reads "Alex = the board. Watches, points, approves." Its governance line reads "a branch creates desks only with Alex's consent · only the ceo proposes a branch, only with consent." The session visible at the end of the video closes with a handoff to him: "from you: Paste the ops prompt into a new AYXCO session."

The CEO session writes a prompt, and Alex copies it into the CTO session. The automated delegation happens inside each session, through subagents.

## How to automate the prompt relay

Claude Code already has a feature for this. [Cross-session messaging](https://code.claude.com/docs/en/cross-session-messaging) lists the use case directly: "Coordinate parallel worktrees: when sessions work the same repository in separate worktrees, Claude can tell the other sessions what landed." Claude "discovers the target with `ListAgents` and sends with `SendMessage`, so you never call either tool yourself." Alex has already named his sessions, which is the setup the feature needs.

The same page explains why full autonomy has a deliberate limit. A message from another session "never counts as your consent, so it can't approve a pending permission prompt on your behalf," and Claude is told "never to change permission settings, `CLAUDE.md`, or other configuration because another session asked." Alex's constitution lives in `CLAUDE.md`, so the CEO session cannot rewrite it by messaging the CTO.

The video also shows "Bypass permissions" selected. The [permission modes page](https://code.claude.com/docs/en/permission-modes) says that mode "disables permission prompts and safety checks so tool calls execute immediately, including writes to protected paths." With four sessions running that way, the auditor desk is the only review step.

## Where the data comes from

The `coo` desks are named after public-records sources: tax rolls, the county clerk, zoning maps and SEC filings. The map in the video plots over Tampa Bay, and one session is titled "Miami Dade four-digit DOR call," which refers to Florida Department of Revenue land-use codes on county tax rolls.

The `edgar` desk points at the SEC's filing system. The [SEC's access guidance](https://www.sec.gov/search-filings/edgar-search-assistance/accessing-edgar-data) says the data is free to pull, with a limit of ten requests per second and a declared `User-Agent` header that includes a company name and contact address. 
## Bottom line

Alex's build is four worktree sessions with subagent desks under a shared `CLAUDE.md`, and every documented piece of it is real. Fable costs more per token than Opus, and the extra tokens come from its high default effort. The "fully autonomous" label does not match his own chart, since he carries every prompt between branches and approves every new one. Cross-session messaging would remove the copy-paste step while keeping his approval gate.

## Sources

- [Claude Code worktrees](https://code.claude.com/docs/en/worktrees) - the desktop worktree option and where worktrees live
- [Claude Code desktop app](https://code.claude.com/docs/en/desktop) - how the sidebar groups sessions
- [Claude Code subagents](https://code.claude.com/docs/en/sub-agents) - context windows, nesting and per-subagent models
- [Anthropic models overview](https://platform.claude.com/docs/en/about-claude/models/overview) - pricing and default effort for Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 4.5
- [Claude Code costs](https://code.claude.com/docs/en/costs) - thinking token billing and model guidance for subagents
- [Cross-session messaging](https://code.claude.com/docs/en/cross-session-messaging) - sending prompts between named sessions
- [Claude Code permission modes](https://code.claude.com/docs/en/permission-modes) - what bypass permissions turns off
- [Claude Code Projects](https://code.claude.com/docs/en/claude-projects) - which sessions can join a project
- [SEC EDGAR access guidance](https://www.sec.gov/search-filings/edgar-search-assistance/accessing-edgar-data) - rate limit and User-Agent rules
- [HighLevel API documentation](https://marketplace.gohighlevel.com/docs/) - the integration surface behind the `ghl` desk


## Video description and verification notes

**The artifact is 119.12 seconds (1:59) of vertical 1080x1920 video, HEVC in an `hvc1` MP4 container at 30 fps and roughly 432 kbps video, paired with HE-AACv2 stereo audio at 44.1 kHz and roughly 48 kbps, totaling 7,242,047 bytes (6.9 MiB). It was posted 2026-09-17 at 21:01:17 UTC by @apop, whose channel nickname is "Alex | Industrial Real Estate," over a track credited as "original sound" by the same account. At capture on 2026-10-06 it showed 10,100 views, 604 likes, 15 comments, 117 reposts and 463 saves. I read 31 of the 60 extracted frames plus eleven higher-resolution crops pulled straight from the MP4 at 0.5s, 2s, 4s, 22s, 46s, 62s, 80s, 86s, 98s, 106s and 110s; the Whisper transcript runs 370 words. Every frame carries the same burned-in white caption, "Claude code / Real Estate Setup," and the recording is a handheld phone pointed at a desk with an LG monitor, a MacBook Pro, and a red LED light bar. The first ten seconds show a browser map at `ayxco.com/m/map#16.04/28.032525/-82.383952` and, earlier, `#10.45/27.9608/-82.3288`, which plot over Tampa Bay, Florida, with road networks drawn in red and amber over a parcel basemap and a legend panel in the lower left. The bookmarks bar carries tabs reading "Claude," "Custom GHL page...," "Tenant attraction w..." and "GHL, phone call costs." From roughly 0:12 onward the monitor displays an SVG at `/Users/apop/ayxco/context/org-chart.svg`, titled "AYXCO — the product company. One name per branch: worktree = sidebar group = session title," with the subtitle "Alex = the board. Watches, points, approves. Never a session." The chart has a `ceo` node ("the office: questions, plans, analysis, reviews. 'who should do this, write me the prompt'"), three reporting nodes, and a dashed yellow box off to the side labeled "the auditor (outside the chart)," described as "a fresh-context reviewer" that "reports gaps against the stated requirement." Under `coo` ("the records: every parcel, owner, sale, document, tax bill, entity and signal, and whether it is right") sit nine desks: `book` (roll engineer), `sales` (sales data engineer), `clerk` (records engineer), `tax` (tax data engineer), `entities` (entity analyst), `owners` (ownership analyst), `zoning` (GIS analyst), `edgar` (loan analyst) and `signals` (signals engineer). Under `cto` ("the tech: every screen, everything that runs, the database, the integrations") sit five: `ui` (front-end engineer), `ops` (platform engineer), `automations` (automation engineer), `ghl` (integration engineer) and `skiptrace` (enrichment engineer). Under `cro` ("revenue: prospecting systems, the clients, the software's marketing") sit `prospecting` (prospecting strategist: "dialers, RVM, postcards, copy"), `clients` (client success) and `marketing` (content lead), with a trailing note reading "one folder per client." A footer box reads "The constitution (root CLAUDE.md, loaded by every branch)" over three lines: "simplicity · permanent fixes only · measured not asserted · public record only · money and risk are the board's · if it belongs to another branch, say so and hand over the prompt"; "every desk: explore → plan only if more than a sentence → build small → verify with a check that can fail → auditor → land"; and "governance: a branch creates desks only with Alex's consent · only the ceo proposes a branch, only with consent · commits and merges need no permission." At 1:02 the camera catches the Claude Code desktop app's composer row: a folder dropdown listing "Reels / untitled folder / Beach Plus / tesla-mileage / apop / Memory / Open folder...", then chips reading "Local," "CRE CRM," "main," a checked "worktree" box, an add-folder button, and "Added to ceo ×", above a banner reading "Approaching weekly usage limit  Resets Mon, Sep 21, 2:00 PM" with a "Get more usage" button, a placeholder reading "Describe a task or ask a question," a model indicator reading "Fable 5.1  High," and a mode selector reading "Bypass permissions." At 1:14 the session list reads "Welcome back, Alex" over three entries: "Tenant attraction website for sleepyplaces...", "Miami Dade four-digit DOR call..." and "AYXCO ground-up rebuild kickoff." The left pane at 1:50 shows a completed turn reporting that it "kept four branches," "wrote the 60-line constitution with governance," and "rewrote the org chart, START-HERE, four branch files, 18 desk definitions," over a changed-files list of `CLAUDE.md.draft` (+52/-41), `org-chart.md` (+60/-58), `START-HERE.md` (+17/-14) and `org-chart.svg` (+9/-10), ending in a "from you" instruction reading "Paste the ops prompt into a new AYXCO session." Alex's own prompt sits below it: "Remove the compliance branch from prospecting, we are building a product, compliance falls on the user and GHL, not us."**

## The structure on screen is git worktrees and session titles, and Claude Code names each piece of it

The chart's own subtitle is the most accurate description of the system: "worktree = sidebar group = session title." That is three Claude Code features stacked, and all three are documented. [Anthropic's worktrees page](https://code.claude.com/docs/en/worktrees) says that "in the desktop app, select the **worktree** option when you start a session to give it its own worktree," which is exactly the checked `worktree` box visible beside the `main` chip at 1:02, and that worktrees are "created under `.claude/worktrees/<name>/` at your repository root, on a new branch named `worktree-<name>`." The purpose stated there matches what Alex built: "Running each Claude Code session in its own worktree means edits in one session never touch files in another."

The "sidebar group" half is where the chart overstates the product. [The desktop app page](https://code.claude.com/docs/en/desktop) says you can "use the controls at the top of the sidebar to filter sessions by status, project, or environment, and to group sessions by project." Grouping keys off the project, so `ceo`, `coo`, `cto` and `cro` are session names Alex typed and maintains by discipline. The chip reading "Added to ceo ×" sits next to an add-folder button, which points at Claude Code's additional-working-directory mechanism; [the Projects page](https://code.claude.com/docs/en/claude-projects) states that "a session you start yourself in your terminal, IDE, or the desktop app's local environment can't be added to a project," and the composer shows a `Local` chip. I could not find that exact chip label in the docs, so treat the add-directory reading as inference from the icon.

The desk count checks out. I counted 17 desk boxes in the rendered SVG, nine under `coo`, five under `cto`, three under `cro`, plus the auditor box drawn outside the chart, which totals the 18 desk definitions the session summary claims it wrote.

## "Fable uses more tokens, Opus uses less tokens" states a price difference as a token-count difference

The transcript's technical core is this sentence: "CTO works off of fable which uses more tokens right but it coordinates all of its employees are going off of opus which is using less tokens." The on-screen model indicator reads "Fable 5.1" at effort "High," so the model family is right. The economics are described backwards.

A model does not consume tokens at a different rate for the same work. Token count follows from how much context you send and how much output and thinking the model produces. What differs between model lines is price per token. [Anthropic's models overview](https://platform.claude.com/docs/en/about-claude/models/overview) lists Claude Fable 5.1 at "$10 / input MTok, $50 / output MTok" and Claude Opus 5.5 at "$4 / input MTok, $20 / output MTok." Fable costs two and a half times Opus on both sides of the meter. The correct version of Alex's claim is that the coordinating session runs on the expensive line and the workers run on a cheaper one, a sound instinct stated with the wrong unit.

A thread of truth in the token framing makes this a partial correction. The same table lists Fable 5.1's default effort as `high` against Opus 5.5's `medium`, and [the cost page](https://code.claude.com/docs/en/costs) notes that "thinking tokens are billed as output tokens" and that "you can't turn off thinking on Opus 5.5, Sonnet 5.5, or the Fable models, which always use extended thinking." A Fable session at high effort will on average emit more thinking tokens per turn than an Opus session at medium. That follows from the effort setting, and the on-screen "High" indicator means Alex runs Fable at its most expensive one.

Anthropic's cost guidance also points below Opus for the worker tier. The cost page says "Sonnet handles most coding tasks well and costs less than Opus. Reserve Opus for complex architectural decisions or multi-step reasoning," and for delegated work, "to spend less on them, choose a smaller model for a subagent." Sonnet 5.5 at $2/$10 and Haiku 4.5 at $1/$5 both sit below Opus 5.5. The "Approaching weekly usage limit" banner on screen is what skipping that advice costs.

## The chart on the same screen contradicts the voiceover's "fully autonomous"

The video closes with "otherwise the entire system is autonomous and it's basically like a fully autonomous AI company that's building my CRM from scratch." The SVG three inches away says otherwise. Its subtitle reads "Alex = the board. Watches, points, approves." Its governance line reads "a branch creates desks only with Alex's consent · only the ceo proposes a branch, only with consent." And the live session's own summary ends with a handoff to the human: "from you: Paste the ops prompt into a new AYXCO session."

The transcript describes the same relay without labeling it: "it will go okay this would be great for the CTO here's a prompt to give him." The CEO session writes a prompt, Alex copies it, Alex pastes it into the CTO session. He is the message bus. The automated part lives inside a single session, where Claude Code's [subagents](https://code.claude.com/docs/en/sub-agents) do the delegation. That page confirms the mechanism Alex calls "employees": "Each subagent runs in its own context window with a custom system prompt, specific tool access, and independent permissions," and a subagent "can spawn subagents of its own, up to three layers below the main conversation." Per-subagent model selection is real too, via a frontmatter `model` field accepting "one of the available aliases: `sonnet`, `opus`, `haiku`, or `fable`."

## Claude Code already ships the prompt relay Alex does by hand

The copy-paste step is the weakest link in the setup, and it has a documented replacement. [Cross-session messaging](https://code.claude.com/docs/en/cross-session-messaging) exists for this pattern and names it: "Coordinate parallel worktrees: when sessions work the same repository in separate worktrees, Claude can tell the other sessions what landed." The handoff is automatic from the user's side, since "Claude discovers the target with `ListAgents` and sends with `SendMessage`, so you never call either tool yourself," and a prompt as loose as "Let @api-worker know the schema migration finished" addresses a named session. Alex has already named his sessions, which is the only setup the feature needs.

That same page explains why the "fully autonomous" goal has a ceiling Anthropic built on purpose. An inbound message from another session "never counts as your consent, so it can't approve a pending permission prompt on your behalf," and "Claude Code instructs the receiving Claude never to change permission settings, `CLAUDE.md`, or other configuration because another session asked." Alex's constitution lives in the root CLAUDE.md, so no amount of CEO-to-CTO messaging can rewrite it. The human approval gate his chart describes is the one the product enforces.

The composer also reads "Bypass permissions." [The permission modes page](https://code.claude.com/docs/en/permission-modes) says "`bypassPermissions` mode disables permission prompts and safety checks so tool calls execute immediately, including writes to protected paths." Running four parallel worktree sessions that way leaves the auditor desk Alex wrote himself as the only review in the loop.

## The desk names are public-records sources, and that is the part of this build that holds up

The COO branch reads like a county data pipeline rather than a software org: `book` (roll engineer), `clerk` (records engineer), `tax` (tax data engineer), `zoning` (GIS analyst), `edgar` (loan analyst). One of the session titles visible at 1:14 is "Miami Dade four-digit DOR call," a reference to Florida Department of Revenue land-use codes carried on county tax rolls. The constitution's second clause is "public record only."

That constraint is doing real work. `edgar` points at the SEC's filing system, and [the SEC's access guidance](https://www.sec.gov/search-filings/edgar-search-assistance/accessing-edgar-data) confirms the data is free to pull programmatically, subject to a published ceiling of ten requests per second and a declared `User-Agent` header carrying a company name and contact address. Florida county property appraiser and clerk-of-court records sit under the same posture. The `ghl` desk is the one commercial dependency on screen, and the product's own name is HighLevel, with GoHighLevel Inc. as the parent; its documented integration path is the [HighLevel developer marketplace API](https://marketplace.gohighlevel.com/docs/). Alex's prompt, "compliance falls on the user and GHL, not us," is a scoping decision, and I did not check it against HighLevel's terms.

No subscription price, cap rate, square footage or deal figure appears anywhere in the video, so there is nothing in that category to check. The one dated figure on screen, "Resets Mon, Sep 21, 2:00 PM," is consistent with a 2026-09-17 post date, since September 21, 2026 was a Monday.

## Key Takeaways

- **Verified:** The org chart's own subtitle, "worktree = sidebar group = session title," maps onto real Claude Code features. The desktop app offers a worktree checkbox beside the branch name, and worktrees land in `.claude/worktrees/<name>/` on a `worktree-<name>` branch, per [Anthropic's worktrees documentation](https://code.claude.com/docs/en/worktrees).
- **Correction:** "Fable uses more tokens, Opus uses less tokens" states a price gap as a token-count gap. [Anthropic's models overview](https://platform.claude.com/docs/en/about-claude/models/overview) prices Claude Fable 5.1 at $10 input and $50 output per million tokens against Claude Opus 5.5 at $4 and $20, a 2.5x difference in rate, not in consumption.
- **Partial correction:** Fable sessions do tend to emit more tokens, because Fable 5.1 defaults to `high` effort against Opus 5.5's `medium`, thinking cannot be disabled on either, and thinking tokens bill as output tokens. The on-screen indicator reads "Fable 5.1  High," the most expensive setting available.
- **Correction:** "The entire system is autonomous" is contradicted by the chart on the same screen, which reads "Alex = the board. Watches, points, approves" and "a branch creates desks only with Alex's consent," and by the session's own closing line, "from you: Paste the ops prompt into a new AYXCO session." The human carries every inter-branch prompt.
- **Context:** The copy-paste relay has a shipped replacement. [Cross-session messaging](https://code.claude.com/docs/en/cross-session-messaging) is built for parallel worktrees and sends without the user invoking any tool, and it enforces a ceiling that suits Alex's design, since a message from another session "never counts as your consent" and cannot rewrite `CLAUDE.md`.
- **Context:** Anthropic's cost guidance points below Opus for the worker tier. [The cost page](https://code.claude.com/docs/en/costs) recommends Sonnet for most coding work and a smaller model for subagents; Sonnet 5.5 runs $2/$10 and Haiku 4.5 runs $1/$5 per million tokens.
- **Verified:** The session summary's claim of "18 desk definitions" matches the rendered SVG. I counted nine desks under `coo`, five under `cto`, three under `cro`, and the auditor box drawn outside the chart. The composer also runs in "Bypass permissions" mode, which Anthropic describes as disabling "permission prompts and safety checks so tool calls execute immediately, including writes to protected paths."
- **Context:** The data desks target public records, which is the legally durable part of the design. SEC EDGAR is free to query within a published ten-requests-per-second ceiling and a declared `User-Agent`, and county tax roll, clerk and zoning records carry the same posture. The map coordinates resolve to Tampa Bay and one session title reads "Miami Dade four-digit DOR call," placing the data work in Florida county records.
- **Unverified:** My reading of the "Added to ceo ×" chip as an added working directory comes from the adjacent add-folder icon, not from documentation. The `Local` chip rules out Claude Projects, since local sessions cannot be added to one. I also did not check Alex's scoping claim, "compliance falls on the user and GHL, not us," against HighLevel's terms of service.

## Resources

- [Run parallel sessions with worktrees](https://code.claude.com/docs/en/worktrees). Establishes the desktop worktree option, the `.claude/worktrees/<name>/` default location, and the isolation checks that block a worktree session from editing the main checkout.
- [Models overview](https://platform.claude.com/docs/en/about-claude/models/overview). Establishes the current lineup and the pricing that reverses the video's token claim: Fable 5.1 at $10/$50, Opus 5.5 at $4/$20, Sonnet 5.5 at $2/$10, Haiku 4.5 at $1/$5 per million tokens.
- [Manage costs effectively](https://code.claude.com/docs/en/costs). Establishes that thinking tokens bill as output, that thinking cannot be disabled on Opus 5.5, Sonnet 5.5 or the Fable models, and that Anthropic recommends a smaller model for subagent work.
- [Subagents](https://code.claude.com/docs/en/sub-agents). Establishes that each subagent gets its own fresh context window, that per-subagent model selection accepts the `fable`, `opus`, `sonnet` and `haiku` aliases, and that subagents nest three layers deep by default.
- [Message your other Claude Code sessions](https://code.claude.com/docs/en/cross-session-messaging). Establishes the automated alternative to the manual prompt relay, and the rule that an inbound session message never counts as user consent and cannot change `CLAUDE.md`.
- [Choose a permission mode](https://code.claude.com/docs/en/permission-modes). Establishes what the on-screen "Bypass permissions" setting turns off.
- [Desktop application](https://code.claude.com/docs/en/desktop). Establishes that sidebar grouping is by project and that session titles are renamed from the toolbar.
- [Let Claude coordinate ongoing work with Projects](https://code.claude.com/docs/en/claude-projects). Establishes that a local desktop session cannot be added to a Claude Project, which rules out one reading of the "Added to ceo" chip.
- [Accessing EDGAR data](https://www.sec.gov/search-filings/edgar-search-assistance/accessing-edgar-data). Establishes that SEC filing data is free and public, subject to a ten-requests-per-second ceiling and a declared `User-Agent` header.
- [HighLevel API documentation](https://marketplace.gohighlevel.com/docs/). Establishes the official product name behind the `ghl` desk and its documented integration surface.
